A Hierarchical Subject-Session Transfer Learning Strategy for Lower-Limb Motor Imagery Brain-Computer Interface in Individuals with Spinal Cord Injury

Abstract

Motor imagery (MI) brain computer interface (BCI) offer a non-invasive avenue for people with spinal cord injury (SCI) to volitionally control assistive technologies. However, MI-BCI performance in SCI is constrained by pronounced inter-subject variability, stemming from differences in injury level, chronicity, and the degree of cortical reorganization induced by sensory deafferentation rather than direct cortical damage. We propose a hierarchical subject-session transfer (HSST) method that first pre-trains on other subjects, then adapts with frozen features on target subject's support sessions, and finally personalizes on the target session using an CNN model backbone. In five SCI subjects (four sessions each), HSST achieved the highest accuracy (89.8±4.6%), surpassing ST (86.8±5.7%), CST (87.0±7.2%), and SST (87.8±7.5%).